SBBTS: A Unified Schrödinger–Bass Framework for Synthetic Financial Time Series

Battle Count: 93

Highly relevant to quantitative trading. The paper directly addresses synthetic data generation for financial time series, which is critical for backtesting, stress testing, and training predictive models. The data augmentation experiments on S&P 500 show concrete improvements in Sharpe ratio (from 1.613 to 2.113) and classification accuracy. The framework captures stochastic volatility and cross-asset correlations essential for realistic market simulation. However, the trading strategy used is a simple toy model without transaction costs, and the primary contribution is the generative framework rather than a trading strategy per se.

The implementation involves multiple components: (1) the SBBTS theoretical framework with conditional transport decomposition, (2) neural network architecture with encoder-only transformer for trajectory embedding, (3) iterative training with Brownian bridge sampling, (4) dimensionality reduction via PCA and clustering for high-dimensional data, and (5) integration with TabICL for downstream forecasting. The mathematical formulation (Schrödinger-Bass Bridge, semimartingale optimal transport) requires advanced knowledge. The code is available but the full pipeline from raw data to synthetic generation to downstream evaluation is non-trivial to reproduce.

Code is publicly available on GitHub (https://github.com/alexouadi/SBBTS). The paper provides detailed algorithm descriptions (Algorithm 1), neural network architecture (Appendix B), hyperparameters (Table 2), and experimental setup details. However, the S&P 500 data source is referenced externally [6] and the exact preprocessing pipeline for TabICL requires careful replication. The theoretical decomposition (Theorem 3.2) is fully proven in the appendix.